The Study of Small Area Estimation Using Oversampling and M-Quantile Robust Regression Approach

Abstract

Statistics Indonesia (BPS) calculates poverty indicators (Head Count Ratio, Poverty Gap, and Poverty Severity) using National Socio-Economic Survey (Susenas). Susenas is only designed to estimate province and municipality/regency area level, whereas the government requires estimation until smaller area level (sub-district and village). Estimating poverty indicators directly from Susenas for the smaller area often leads to inaccurate estimates. To solve this problem, BPS usually conduct additional survey called Regional Socio-Economic Survey (Suseda) by increasing number to the original sample (called oversampling) but with the very high cost. Therefore, we proposed small area estimation technique which based on the unit level model using Population Census 2010 (SP2010) as the population auxiliary variables and household per-capita expenditure (Susenas 2015) as the response variable. We utilized robust M-quantile regression model which robust to the outlier using three weight functions (Huber, Hampel, and Tukey Bisquare). Our results provide evidence that M-quantile model is more accurate than direct estimates with oversampling.